OpenLake launches at YC with inference that keeps training the model

Most companies that want a model tuned to their own data run into the same problem. They need compute, training infrastructure, evaluation and a team to run it. OpenLake launched at Y Combinator Fall 2026 to remove that step. The company's launch page on Y Combinator has 49 upvotes and tags for Generative AI, SaaS, Big Data and Infrastructure.

OpenLake picks an open model, keeps training it on new company data, and serves the new version in place.

What the service does

A team chooses an open model, such as DeepSeek V4 or GLM 5.3, and deploys it on OpenLake. The team plugs in its datasets, reinforcement-learning environments and data sources. OpenLake runs the post-training as new data arrives. The serving layer swaps in the new version. Each new version can be A/B tested against the ones before it.

The company says the trained checkpoints belong to the customer. A team can export and download them. OpenLake says the product aims at better performance on the tasks the customer cares about. Public benchmark scores are not the goal.

OpenLake says its models cut serving costs in half on average. That is the company's own figure. No outside team has repeated it.

The launch page frames the problem this way. Foundation models are trained on the public world. A company's own data, its people and its customers' problems are never seen by a frontier model. Fine-tuning a model on that data today means building an in-house research lab.

About the founder

The launch post is written by Arnav Balyan. He says he worked on core data and AI systems at Uber before OpenLake. He says Uber products that used those systems reached millions of people. The product is built on the idea that models are not the hard part. Keeping them current is.

The company is OpenLake Technologies, in the Y Combinator Fall 2026 batch, listed in B2B. The launch page carries a video and a demo link, but no customer or price.

What we can and cannot say

OpenLake's launch page names no customer. The post has no round amount, no public price list and no named outside backer. The "cut serving costs in half" figure is OpenLake's own claim. There is no outside benchmark that a reader can check.

The product is available today at openlakehq.com. The launch post asks teams running models in production to book a call. A buyer should ask OpenLake for a worked example on a task the team already understands. Until that exists, 49 YC votes and one founder from Uber's data team are the signal.

Source

• Primary source: Y Combinator launch: OpenLake, by Arnav Balyan.
• Company site: openlakehq.com.